Didrien/meteorix-rfdetr-small

Meteorix RF-DETR Small (meteor detection)

RF-DETR Small fine-tuned from the COCO-pretrained rf-detr-small.pth checkpoint to detect meteors (single class meteor) in ground-based night-sky images, as part of the Meteorix study comparing YOLO26n, RF-DETR Small and DETR-ResNet-50.

Usage

from huggingface_hub import hf_hub_download
from PIL import Image
from rfdetr import RFDETRSmall

model = RFDETRSmall(pretrain_weights=hf_hub_download("Didrien/meteorix-rfdetr-small", "meteorix-rfdetr-small.pth"), num_classes=1)
image = Image.open("sky.jpg").convert("RGB")  # grayscale images must be converted to RGB
detections = model.predict(image, threshold=0.5)
boxes = detections.xyxy  # pixel coordinates in the original image

Tested with rfdetr==1.6.5.post0. The model works at the default RF-DETR Small resolution (512x512).

Training

• Default rfdetr fine-tuning recipe: AdamW (learning rate 1e-4, layer-wise decay inside the backbone), EMA of the weights, effective batch size 16 (4 x 4 gradient accumulation), no learning-rate decay, 100 epochs.
• The weights are the checkpoint_best_total.pth selected by rfdetr on the validation split: EMA weights of epoch 25 (counted from 0, i.e. the 26th epoch; validation mAP50-95 0.670).
• Data: Mendeley meteor image dataset (Gradvohl & Shirasuna, DOI 10.17632/4k98n84d9g.1, CC BY 4.0), 3,849 training / 550 validation / 1,100 test images.

Results (image level, confidence 0.5, matching IoU 0.5)

Test set	Precision	Recall	F1
Mendeley test split (1,100 images)	0.985	0.971	0.978
NEO-Bench UAEMMN, all splits (1,382 images)	0.976	0.772	0.862

Mean end-to-end latency on an RTX 4090 (image loading, pre-processing, forward pass, post-processing): 18.4 ms per image.